We also see improvements like 7x between the current and last Google TPU generations, with different chips for training vs inference.